What changed

MLCommons published MLPerf Training 6.1 with 153 systems and 250 results from 24 organizations. The round adds a Qwen3.5-397B GRPO post-training workload and a new HSTU recommender benchmark, and includes first submissions using NVIDIA Rubin-generation GPUs and Google TPUv7 Ironwood.

Why it matters

An open, peer-reviewed benchmark with multiple vendors gives infrastructure buyers a more comparable view than single-vendor launch claims. Adding post-training makes the suite more relevant to agent systems whose economics extend beyond pretraining and inference.

What to do

Filter the public result table to the exact workload, scale, availability class and power or software constraints that match your deployment. Shortlist comparable systems, then reproduce throughput, convergence, reliability and cost on private data before buying capacity.

Caveat

MLPerf controls the benchmark but submitters configure and optimize their own systems. Results do not include procurement price or application quality, and some new hardware or endpoint results can be provisional or not yet generally available.